Related Experiment Videos
Privacy-Preserving Federated Deep Learning for Robust Anomaly Detection in Distributed Security Sensing Systems.
Di Xu1,2, Hongli Chen2,3, Yansen Zeng1,2
1China Agricultural University, Beijing 100083, China.
Sensors (Basel, Switzerland)
|June 26, 2026
Summary
A new federated deep anomaly detection framework enhances financial security by enabling collaborative modeling without sharing sensitive data. This approach ensures robust monitoring of transactions and systems, protecting privacy in distributed financial networks.
Area of Science:
- Financial Technology (FinTech)
- Cybersecurity
- Artificial Intelligence (AI)
- Machine Learning (ML)
Background:
- Financial security sensing data is multisource heterogeneous, temporally dynamic, and privacy-sensitive.
- Traditional centralized anomaly detection fails to meet requirements for cross-institutional collaboration, privacy protection, and robust monitoring.
- Distributed financial systems necessitate novel approaches for anomaly detection.
Purpose of the Study:
- To propose a data-local federated deep anomaly detection framework for distributed financial security sensing systems.
- To enable cross-node collaborative modeling while protecting client data privacy.
- To achieve robust monitoring of transaction and system anomalies in heterogeneous financial data.
Main Methods:
- Constructed a local deep financial security sensing representation module for temporal encoding and attention-based modeling.
- Developed a data-local federated optimization and personalized aggregation mechanism for cross-node knowledge sharing.
- Employed local personalized detection heads for non-IID data and an adversarially robust strategy for enhanced stability.
Main Results:
- Achieved high performance in financial anomaly detection: 92.37% Accuracy, 89.41% Precision, 88.26% Recall, 88.83% F1-score, 93.06% AUC.
- Demonstrated superior performance against baseline methods including Isolation Forest, Autoencoder, LSTM, Transformer, FedAvg, FedProx, SCAFFOLD, and MOON.
- Showcased robustness against various perturbations and malicious client scenarios, maintaining high F1-scores.
Conclusions:
- The proposed data-local federated deep anomaly detection framework effectively addresses challenges in distributed financial security.
- The framework ensures data privacy, handles non-IID data, and provides robust anomaly detection capabilities.
- This approach offers an efficient intelligent solution for financial AI security monitoring with data localization and heterogeneity.